Localizing Multiple Bugs in RTL Designs by Classifying Hit-Statements Using Neural Networks

Abstract

Nowadays the advanced applications required in our lives have led to a significant increase in the complexity of circuits, which enhances the possibility of occurring design errors. Hence an automated, powerful, and scalable debugging approach is needed. Therefore, this paper proposes a scalable approach for localizing multiple bugs in Register-Transfer level (RTL) designs by using neural networks. The main idea is that hit-statements which are covered by failed test-vectors are more suspicious than those covered by passed test-vectors. We use coverage data as samples of our data set, label these samples, and tune the neural network model. Then we encode hit-statements and give them to the tuned model as new samples. The model classifies hit-statements. Hit-statements that take the failed labels, labels related to the failed test-vectors, are more suspicious of containing bugs. The results demonstrate that the proposed methodology outperforms recent approaches Tarsel and CirFix by localizing 80% of bugs at Top-1. The results also imply that our methodology increases the F₁-score metric by 1.13× in comparison with existing RTL debugging techniques, which are prediction-based.